Data driven subword unit modeling for speech recognition and its application to interactive reading tutors

This paper proposes a novel token-passing search architecture for supporting subword unit based speech recognition and a corresponding algorithm based on the well-known LZW text compression method to determine a vocabulary of subword units in an unsupervised manner. We compare our subword unit selection algorithm to an existing approach based on Minimum Description Length (MDL) modeling and also syllable representations for English. Our approach is shown to offer units which share properties similar to syllables, but are determined in a language-independent and data-driven manner. Using our novel token passing architecture which combines both word-level and subword unit representations, we applied the proposed framework to the problem of oral reading tracking within an interactive literacy tutor for children. The proposed architecture is shown to provide advantages over whole-word based speech recognition for the problem of recognizing and detecting oral reading events.